"It became a self-fulfilling prophecy": How Lived Experiences are Entangled with AI Predictions in Menstrual Cycle Tracking Apps

Technical Summary: "It became a self-fulfilling prophecy": How Lived Experiences are Entangled with AI Predictions in Menstrual Cycle Tracking Apps

Problem Statement
Menstrual Cycle Tracking Applications (MCTAs) increasingly utilize AI-based features to provide personalized predictions regarding menstrual cycles, fertility, mood, and physical symptoms (e.g., cramps, bloating, anxiety). While these tools aim to enhance self-awareness and health management, they suffer from a "black-box" nature where the generation of predictions lacks transparency. Existing research indicates that AI in well-being technologies often simplifies complex human experiences, leading to mismatches between algorithmic outputs and users' lived realities. Crucially, there is a gap in understanding how users interact with these predictive features, particularly regarding subjective states like mood where the "ground truth" is known only to the user. This opacity risks fostering over-reliance, where users may trust algorithmic predictions over their own bodily sensations, potentially creating a "self-fulfilling prophecy" where the prediction influences the experience itself. Furthermore, current Explainable AI (XAI) approaches often fail to provide usable or effective explanations in commercial well-being contexts, leaving users without the resources to critically engage with these entangled interactions.

Methodology
The authors employ a mixed-method qualitative approach consisting of two distinct studies to investigate human-AI entanglement in MCTAs:

1. Study 1: Semi-Structured User Interviews

   • Participants: 14 female participants (aged 18–35) from Northern and Central Europe, comprising both current and former MCTA users with varying cycle types (regular, irregular, absent).
   • Procedure: Interviews focused on participants' understanding of AI features, their logging practices, and their lived experiences with AI predictions and explanations. Participants were shown screenshots of the "Flo" app (a market leader) to discuss specific AI features, including cycle predictions, symptom insights, and fertility forecasts.
   • Analysis: Data was analyzed using Braun and Clarke's six-phase framework for reflexive thematic analysis, focusing on how users interpret and co-construct meaning with AI.

2. Study 2: Group Autoethnography

   • Participants: The four authors of the paper, representing diverse bodily conditions: regular cycles, irregular cycles (PCOS), perimenopause, and menopause (on hormone replacement therapy).
   • Procedure: Over an eight-week period, authors tracked moods and symptoms in various MCTAs (Flo, Clue, Stardust, Balance) while maintaining personal journals and capturing interface artifacts (screenshots).
   • Focus: This study aimed to capture first-person, momentary reflections on how interface designs, logging mechanisms, and AI predictions actively shaped their subjective experiences in real-time.
   • Analysis: Individual reflections were synthesized and discussed collaboratively to identify themes regarding the entanglement of logging, prediction, and lived experience, specifically highlighting the experiences of non-normative bodies.

Key Contributions
The paper makes four primary contributions to the HCI and design literature:

1. Elucidating Human-AI Entanglement: It reveals that users' lived experiences and self-knowledge are not merely affected by MCTAs but are co-constituted through the interplay of logging practices, interface designs, predictive AI, and AI explanations.
2. Identifying Resource Gaps: The study demonstrates that users lack adequate resources (such as transparent AI explanations and feedback mechanisms) to establish critical awareness of this entanglement, leading to intuitive but often uncritical reliance on predictions.
3. Highlighting Marginalization: It documents the isolating experiences of non-normative users (e.g., those with PCOS, irregular cycles, or menopause), whose bodies are often pathologized or excluded by the normative assumptions embedded in these systems.
4. Design Implications: The authors propose design strategies that move beyond simple transparency to actively support user autonomy and critical engagement with AI entanglement.

Results
The findings are organized around the concept of "entanglement," where the boundary between the user's internal state and the app's output becomes blurred.

• Intuitive Understanding vs. Lack of Transparency: Users possess an intuitive grasp of how AI predictions work (e.g., understanding that more data leads to better accuracy), yet they lack specific knowledge about the underlying mechanisms. This uncertainty leads to a "guessing" game where users often assume the technology is accurate without verification.
• Co-Construction of Experience (The Self-Fulfilling Prophecy): A central finding is that AI predictions actively shape users' lived experiences. Participants reported that receiving a prediction (e.g., "you may feel anxious today") could prime them to feel that way, or lead them to interpret ambiguous sensations through the lens of the prediction. This phenomenon was described by users as a "self-fulfilling prophecy" or akin to a "horoscope."
• Validation and Over-Reliance: Users frequently rely on AI predictions to validate negative lived experiences (e.g., confirming that feeling sad is "normal" for their cycle phase). When predictions mismatch reality, users often engage in self-reflection or even alter their past logging data to force alignment with the algorithm, rather than questioning the AI's accuracy.
• The Role of Interface Design: The autoethnography revealed that the interface itself (icons, categorical labels, suggested logs) guides how users interpret and report their feelings. For instance, an app suggesting "obsessive thoughts" instead of "anxiety" could reframe a user's experience, creating a feedback loop where the prediction influences the logging, which in turn influences future predictions.
• Exclusion of Non-Normative Bodies: Users with irregular cycles, PCOS, or menopause reported feeling marginalized. The apps' "regularity language" (e.g., labeling a cycle as "late") framed their natural variations as errors or faults. In some cases, the lack of relevant predictions or explanations for non-normative stages led to a sense of isolation and a feeling that the technology was not designed for them.

Significance and Claims
The paper argues that human-AI entanglement in MCTAs is not a negative phenomenon to be eliminated, but a reality to be designed for. The authors claim that current design paradigms, which prioritize accuracy and efficiency, fail to account for the co-constructive nature of AI and lived experience.

The significance of this work lies in shifting the design focus from merely improving prediction accuracy to acknowledging the relationship between the user and the AI. The authors assert that because users' self-knowledge is deeply entangled with AI predictions, "inaccurate" predictions can become the user's reality. Therefore, design interventions should not just aim to correct "errors" but should support users in maintaining autonomy over their meaning-making processes.

The paper concludes by calling for:

• Cognitive Forcing: Designing interfaces that encourage users to log their experiences before seeing predictions to prevent priming effects.
• Transparent Uncertainty: Providing AI explanations that explicitly surface the limitations and uncertainties of predictions, rather than presenting them as facts.
• Inclusive Design: Creating systems that accommodate non-normative bodies by offering options to opt-out of predictions or providing explanations tailored to diverse reproductive stages, thereby moving away from the "normative" assumptions that currently dominate MCTA design.